A SMART IRRIGATION SYSTEM USING THE INTERNET OF THINGS AND MACHINE LEARNING FOR WATER-EFFICIENT CROP MANAGEMENT

Abstract

Irrigation accounts for the largest share of freshwater use in Pakistan, yet in most fields it is still scheduled manually. The result is two opposite failures: over-irrigation, which wastes water, leaches nutrients and displaces air from the root zone, and under-irrigation, which depresses yield before any visible sign of stress appears. This paper presents a smart irrigation system that replaces fixed scheduling with a decision derived from measured field conditions. An Arduino UNO node instrumented with temperature and humidity, soil moisture, rain and light sensors collects field readings and actuates a submersible pump through a relay module. The readings were preprocessed into a labelled dataset described by seven agronomic and meteorological features, namely nitrogen, phosphorus and potassium content, temperature, humidity, pH and rainfall, and four supervised classifiers were trained on a 70% partition and evaluated on the held-out 30%. The random forest achieved the highest accuracy at 99.33%, ahead of the support vector machine at 98.66% and the decision tree at 97.65%, whereas logistic regression reached only 78.33%, indicating that the boundary separating the irrigation classes is not linear in this feature space. The support vector machine nevertheless returned the most balanced precision and recall, at 98.36% and 98.77%, and is therefore the preferred model for deployment, since a false negative withholds water from a crop that needs it. The results show that a commodity sensing node coupled with a non-linear classifier can automate the irrigation decision at a cost compatible with smallholder farming.